diff --git a/examples/cuda/cudaWriteRead.cu b/examples/cuda/cudaWriteRead.cu
index e83754ac7c..433dd58149 100644
--- a/examples/cuda/cudaWriteRead.cu
+++ b/examples/cuda/cudaWriteRead.cu
@@ -26,6 +26,7 @@ int BPWrite(const std::string fname, const size_t N, int nSteps){
   // Set up the ADIOS structures
   adios2::ADIOS adios;
   adios2::IO io = adios.DeclareIO("WriteIO");
+  io.SetEngine("BP5");
 
   // Declare an array for the ADIOS data of size (NumOfProcesses * N)
   const adios2::Dims shape{static_cast<size_t>(N)};
@@ -61,28 +62,23 @@ int BPRead(const std::string fname, const size_t N, int nSteps){
   // Create ADIOS structures
   adios2::ADIOS adios;
   adios2::IO io = adios.DeclareIO("ReadIO");
+  io.SetEngine("BP5");
 
   adios2::Engine bpReader = io.Open(fname, adios2::Mode::Read);
 
-  auto data = io.InquireVariable<float>("data");
-  std::cout << "Steps expected by the reader: " << bpReader.Steps() << std::endl;
-  std::cout << "Expecting data per step: " << data.Shape()[0];
-  std::cout  << " elements" << std::endl;
-
-  int write_step = bpReader.Steps();
-  // Create the local buffer and initialize the access point in the ADIOS file
-  std::vector<float> simData(N); //set size to N
-  const adios2::Dims start{0};
-  const adios2::Dims count{N};
-  const adios2::Box<adios2::Dims> sel(start, count);
-  data.SetSelection(sel);
-  
-  // Read the data in each of the ADIOS steps
-  for (size_t step = 0; step < write_step; step++)
+  unsigned int step = 0;
+  for (; bpReader.BeginStep() == adios2::StepStatus::OK; ++step)
   {
-      data.SetStepSelection({step, 1});
+	  auto data = io.InquireVariable<float>("data");
+	  // Create the local buffer and initialize the access point in the ADIOS file
+	  std::vector<float> simData(N); //set size to N
+	  const adios2::Dims start{0};
+	  const adios2::Dims count{N};
+	  const adios2::Box<adios2::Dims> sel(start, count);
+	  data.SetSelection(sel);
+
       bpReader.Get(data, simData.data());
-      bpReader.PerformGets();
+      bpReader.EndStep();
       std::cout << "Simualation step " << step << " : ";
       std::cout << simData.size() << " elements: " << simData[1] << std::endl;
   }
diff --git a/source/adios2/engine/bp5/BP5Writer.tcc b/source/adios2/engine/bp5/BP5Writer.tcc
index af19b388d8..d563781fde 100644
--- a/source/adios2/engine/bp5/BP5Writer.tcc
+++ b/source/adios2/engine/bp5/BP5Writer.tcc
@@ -26,6 +26,10 @@ void BP5Writer::PutCommon(Variable<T> &variable, const T *values, bool sync)
         BeginStep(StepMode::Update);
     }
     variable.SetData(values);
+    // if the user buffer is allocated on the GPU always use sync mode
+    bool isCudaBuffer = (variable.m_MemorySpace == MemorySpace::CUDA);
+    if (isCudaBuffer)
+        sync = true;
 
     size_t *Shape = NULL;
     size_t *Start = NULL;
diff --git a/source/adios2/helper/adiosMemory.h b/source/adios2/helper/adiosMemory.h
index a67abc904f..d197876d61 100644
--- a/source/adios2/helper/adiosMemory.h
+++ b/source/adios2/helper/adiosMemory.h
@@ -46,6 +46,9 @@ void InsertToBuffer(std::vector<char> &buffer, const T *source,
 template <class T>
 void CopyFromGPUToBuffer(std::vector<char> &buffer, size_t &position,
                          const T *source, const size_t elements = 1) noexcept;
+template <class T>
+void CudaMemCopyToBuffer(char *buffer, size_t position, const T *source,
+                         const size_t size) noexcept;
 
 /**
  * Wrapper around cudaMemcpy needed for isolating CUDA interface dependency
diff --git a/source/adios2/helper/adiosMemory.inl b/source/adios2/helper/adiosMemory.inl
index cfc15ef971..efdb878be0 100644
--- a/source/adios2/helper/adiosMemory.inl
+++ b/source/adios2/helper/adiosMemory.inl
@@ -79,10 +79,17 @@ template <class T>
 void CopyFromGPUToBuffer(std::vector<char> &buffer, size_t &position,
                          const T *source, const size_t elements) noexcept
 {
-    const char *src = reinterpret_cast<const char *>(source);
-    MemcpyGPUToBuffer(buffer.data() + position, src, elements * sizeof(T));
+	CudaMemCopyToBuffer(buffer.data(), position, source, elements * sizeof(T));
     position += elements * sizeof(T);
 }
+
+template <class T>
+void CudaMemCopyToBuffer(char *buffer, size_t position,
+                         const T *source, const size_t size) noexcept
+{
+    const char *src = reinterpret_cast<const char *>(source);
+    MemcpyGPUToBuffer(buffer + position, src, size);
+}
 #endif
 
 template <class T>
diff --git a/source/adios2/toolkit/format/bp5/BP5Serializer.cpp b/source/adios2/toolkit/format/bp5/BP5Serializer.cpp
index 019f1a8cea..0902be3bf8 100644
--- a/source/adios2/toolkit/format/bp5/BP5Serializer.cpp
+++ b/source/adios2/toolkit/format/bp5/BP5Serializer.cpp
@@ -570,15 +570,27 @@ void BP5Serializer::DumpDeferredBlocks(bool forceCopyDeferred)
 }
 
 static void GetMinMax(const void *Data, size_t ElemCount, const DataType Type,
-                      core::Engine::MinMaxStruct &MinMax)
+                      core::Engine::MinMaxStruct &MinMax, MemorySpace MemSpace)
 {
-
     MinMax.Init(Type);
     if (ElemCount == 0)
         return;
     if (Type == DataType::Compound)
     {
     }
+#ifdef ADIOS2_HAVE_CUDA
+#define pertype(T, N)                                                          \
+    else if (MemSpace == MemorySpace::CUDA &&                                  \
+             Type == helper::GetDataType<T>())                                 \
+    {                                                                          \
+        const size_t size = ElemCount * sizeof(T);                             \
+        const T *values = (const T *)Data;                                     \
+        helper::CUDAMinMax(values, ElemCount, MinMax.MinUnion.field_##N,       \
+                           MinMax.MaxUnion.field_##N);                         \
+    }
+    ADIOS2_FOREACH_MINMAX_STDTYPE_2ARGS(pertype)
+#undef pertype
+#endif
 #define pertype(T, N)                                                          \
     else if (Type == helper::GetDataType<T>())                                 \
     {                                                                          \
@@ -669,7 +681,8 @@ void BP5Serializer::Marshal(void *Variable, const char *Name,
         MinMax.Init(Type);
         if ((m_StatsLevel > 0) && !Span)
         {
-            GetMinMax(Data, ElemCount, (DataType)Rec->Type, MinMax);
+            GetMinMax(Data, ElemCount, (DataType)Rec->Type, MinMax,
+                      VB->m_MemorySpace);
         }
 
         if (Rec->OperatorType)
@@ -698,9 +711,10 @@ void BP5Serializer::Marshal(void *Variable, const char *Name,
         {
             if (!DeferAddToVec)
             {
-                DataOffset = m_PriorDataBufferSizeTotal +
-                             CurDataBuffer->AddToVec(ElemCount * ElemSize, Data,
-                                                     ElemSize, Sync);
+                DataOffset =
+                    m_PriorDataBufferSizeTotal +
+                    CurDataBuffer->AddToVec(ElemCount * ElemSize, Data,
+                                            ElemSize, Sync, VB->m_MemorySpace);
             }
         }
         else
diff --git a/source/adios2/toolkit/format/buffer/BufferV.h b/source/adios2/toolkit/format/buffer/BufferV.h
index b55a9dabba..be1d28be80 100644
--- a/source/adios2/toolkit/format/buffer/BufferV.h
+++ b/source/adios2/toolkit/format/buffer/BufferV.h
@@ -41,7 +41,8 @@ class BufferV
     virtual void Reset();
 
     virtual size_t AddToVec(const size_t size, const void *buf, size_t align,
-                            bool CopyReqd) = 0;
+                            bool CopyReqd,
+                            MemorySpace MemSpace = MemorySpace::Host) = 0;
 
     struct BufferPos
     {
diff --git a/source/adios2/toolkit/format/buffer/chunk/ChunkV.cpp b/source/adios2/toolkit/format/buffer/chunk/ChunkV.cpp
index bf1e3a0131..94b14adaaf 100644
--- a/source/adios2/toolkit/format/buffer/chunk/ChunkV.cpp
+++ b/source/adios2/toolkit/format/buffer/chunk/ChunkV.cpp
@@ -7,6 +7,7 @@
  */
 
 #include "ChunkV.h"
+#include "adios2/helper/adiosFunctions.h"
 #include "adios2/toolkit/format/buffer/BufferV.h"
 
 #include <algorithm>
@@ -82,7 +83,7 @@ void ChunkV::CopyExternalToInternal()
 }
 
 size_t ChunkV::AddToVec(const size_t size, const void *buf, size_t align,
-                        bool CopyReqd)
+                        bool CopyReqd, MemorySpace MemSpace)
 {
     if (size == 0)
     {
@@ -120,7 +121,7 @@ size_t ChunkV::AddToVec(const size_t size, const void *buf, size_t align,
         if (AppendPossible)
         {
             // We can use current chunk, just append the data;
-            memcpy(m_TailChunk + m_TailChunkPos, buf, size);
+            CopyDataToBuffer(size, buf, m_TailChunkPos, MemSpace);
             DataV.back().Size += size;
             m_TailChunkPos += size;
         }
@@ -132,7 +133,7 @@ size_t ChunkV::AddToVec(const size_t size, const void *buf, size_t align,
                 NewSize = size;
             m_TailChunk = (char *)malloc(NewSize);
             m_Chunks.push_back(m_TailChunk);
-            memcpy(m_TailChunk, buf, size);
+            CopyDataToBuffer(size, buf, 0, MemSpace);
             m_TailChunkPos = size;
             VecEntry entry = {false, m_TailChunk, 0, size};
             DataV.push_back(entry);
@@ -142,6 +143,19 @@ size_t ChunkV::AddToVec(const size_t size, const void *buf, size_t align,
     return retOffset;
 }
 
+void ChunkV::CopyDataToBuffer(const size_t size, const void *buf, size_t pos,
+                              MemorySpace MemSpace)
+{
+#ifdef ADIOS2_HAVE_CUDA
+    if (MemSpace == MemorySpace::CUDA)
+    {
+        helper::CudaMemCopyToBuffer(m_TailChunk, pos, buf, size);
+        return;
+    }
+#endif
+    memcpy(m_TailChunk + pos, buf, size);
+}
+
 BufferV::BufferPos ChunkV::Allocate(const size_t size, size_t align)
 {
     if (size == 0)
diff --git a/source/adios2/toolkit/format/buffer/chunk/ChunkV.h b/source/adios2/toolkit/format/buffer/chunk/ChunkV.h
index 4db67aed8a..4642108c5f 100644
--- a/source/adios2/toolkit/format/buffer/chunk/ChunkV.h
+++ b/source/adios2/toolkit/format/buffer/chunk/ChunkV.h
@@ -32,7 +32,8 @@ class ChunkV : public BufferV
     virtual std::vector<core::iovec> DataVec() noexcept;
 
     virtual size_t AddToVec(const size_t size, const void *buf, size_t align,
-                            bool CopyReqd);
+                            bool CopyReqd,
+                            MemorySpace MemSpace = MemorySpace::Host);
 
     virtual BufferPos Allocate(const size_t size, size_t align);
     virtual void DownsizeLastAlloc(const size_t oldSize, const size_t newSize);
@@ -40,6 +41,8 @@ class ChunkV : public BufferV
     virtual void *GetPtr(int bufferIdx, size_t posInBuffer);
 
     void CopyExternalToInternal();
+    void CopyDataToBuffer(const size_t size, const void *buf, size_t pos,
+                          MemorySpace MemSpace);
 
 private:
     std::vector<char *> m_Chunks;
diff --git a/source/adios2/toolkit/format/buffer/malloc/MallocV.cpp b/source/adios2/toolkit/format/buffer/malloc/MallocV.cpp
index 7f0015a62f..c23cb1e415 100644
--- a/source/adios2/toolkit/format/buffer/malloc/MallocV.cpp
+++ b/source/adios2/toolkit/format/buffer/malloc/MallocV.cpp
@@ -85,7 +85,7 @@ void MallocV::CopyExternalToInternal()
 }
 
 size_t MallocV::AddToVec(const size_t size, const void *buf, size_t align,
-                         bool CopyReqd)
+                         bool CopyReqd, MemorySpace MemSpace)
 {
     if (size == 0)
     {
diff --git a/source/adios2/toolkit/format/buffer/malloc/MallocV.h b/source/adios2/toolkit/format/buffer/malloc/MallocV.h
index ce1d43ccfc..e78b69b945 100644
--- a/source/adios2/toolkit/format/buffer/malloc/MallocV.h
+++ b/source/adios2/toolkit/format/buffer/malloc/MallocV.h
@@ -36,7 +36,8 @@ class MallocV : public BufferV
     virtual void Reset();
 
     virtual size_t AddToVec(const size_t size, const void *buf, size_t align,
-                            bool CopyReqd);
+                            bool CopyReqd,
+                            MemorySpace MemSpace = MemorySpace::Host);
 
     virtual BufferPos Allocate(const size_t size, size_t align);
     void DownsizeLastAlloc(const size_t oldSize, const size_t newSize);